Research On The Application Of Speech Recognition Technology In Railway Train Operation Simulation Training System

The improvement of my country’s comprehensive strength has promoted the rapid development of domestic railway transportation.The safety of railway transportation has become an issue that cannot be ignored.When the risk factor is minimized within the controllable range of manpower,the traffic commander plays a decisive role.In addition to the shortage of personnel in key locations such as dispatching,the General Manager of Railways also issued documents emphasizing that railway employees must hold certificates to work,and new employees who arrive on duty cannot quickly start the business in a short time,so the station must be equipped with a corresponding training system.With the rapid development of modern technology,the railway train management simulation training system has also explored the combination of voice recognition technology to train technical personnel with high professional quality.However,the current training is still at the level of simple process operation and cannot achieve all-round three-dimensionality.The purpose of training talents.In response to the above problems,this article uses the railway vehicle simulation training system to conduct voice recognition research on railway terminology,and combines cutting-edge technology with traditional industries to improve the training effect and lay the foundation for the realization of intelligent transportation in the future.The main research contents of this paper are as follows:(1)Noise interference in real life will reduce the accuracy of speech recognition.The traditional single-channel noise reduction method has certain limitations for non-stationary speech signals,while the noise reduction method based on the joint dictionary has feature selection and interpretability.The characteristics of,are widely used in speech signal noise reduction,but it still has problems such as insufficient discrimination.Aiming at this problem,an improved joint dictionary noise reduction algorithm is proposed.First,the sparse representation is used to represent the frequency domain amplitude spectrum obtained by the short-time Fourier transform of the speech signal and the noise signal.Second,the sparse representation dictionary is implemented with double sparse representation.Improve dictionary resilience and adaptability,and then add fisher discriminative constraints on this basis,so that the divergence between the speech dictionary and the noise dictionary increases,and the intra-class divergence decreases to increase the discrimination,and finally combined with noise reduction speech Then,the phase spectrum of the signal is recovered by short-time inverse Fourier transform to recover the pure speech signal,so as to achieve the purpose of noise reduction.Experimental results show that the improved joint dictionary noise reduction algorithm proposed in this paper has higher perceptual evaluation of speech quality(PESQ)and output signal-to-noise ratio(SNR)than the other two comparison methods.The effect has been further improved.(2)Domestic speech recognition technology has become mature,but it is mainly for the masses and has not been studied for the railway industry.If the mainstream language recognition software at this stage is simply introduced,the expected recognition effect cannot be achieved.Aiming at the characteristics of railway terminology,a dictionary and training data specific to railway speech are established to explore the acoustic model suitable for railway speech recognition.First,a hidden Markov model is used to combine deep neural networks,cyclic neural networks,and long and short-term memory.The neural network and the two-way long and short-term memory neural network establish four acoustic models.After training,they are applied to the recognition of railway terminology.The experimental results are compared and analyzed.Among the four acoustic models,the Bi LSTM-HMM model is effective in railway terminology speech recognition.The word error rate is the lowest.However,the HMM model itself has certain limitations.Before training the acoustic model,you need to know the corresponding label for the voice alignment process,which requires a certain amount of time and effort.To further optimize the model,this article introduces the sequential chain link model to replace the HMM model.The experimental results show that the Bi LSTM-CTC model is more convenient than the Bi LSTM-HMM hybrid model,and the word error rate of the model is further reduced.